660 likes | 673 Views
Explore the requirements, topologies, and issues related to interconnect and routing in computer architecture. Learn about different routing styles and their impact on throughput and predictability.
E N D
CS184b:Computer Architecture(Abstractions and Optimizations) Day 4: April 4, 2005 Interconnect
Previously • CS184a • interconnect needs and requirements • basic topology • Mostly thought about static/offline routing
This Quarter • This quarter • parallel systems require • typically dynamic switching • interfacing issues • model, hardware, software
Today • Issues • Topology/locality/scaling • (some review) • Styles • from static • to online, packet, wormhole • Online routing
Old Bandwidth aggregate, per endpoint local contention and hotspots Latency Cost (scaling) locality New Arbitration conflict resolution deadlock Routing (quality vs. complexity) Ordering (of messages) Issues
Topology and Locality (Partially) Review
$ $ $ $ P P P P Memory Simple Topologies: Bus • Single Bus • simple, cheap • low bandwidth • not scale with PEs • typically online arbitration • can be offline scheduled
Offline: divide time into N slots assign positions to various communications run modulo N w/ each consumer/producer send/receiving on time slot e.g. 1: A->B 2: C->D 3: A->C 4: A->B 5: C->B 6: D->A 7: D->B 8: A->D Bus Routing
Online: request bus wait for acknowledge Priority based: give to highest priority which requests consider ordering Goti = Wanti ^ Availi Availi+1=Availi ^ /Wanti Solve arbitration in log time using parallel prefix For fairness start priority at different node use cyclic parallel prefix deal with variable starting point Bus Routing
want got avail Priority Arbitration Logic
$ $ $ $ P P P P Memory Token Ring • On bus • delay of cycle goes as N • can’t avoid, even if talking to nearest neighbor • Token ring • pipeline bus data transit (ring) • high frequency • can exit early if local • use token to arbitrate use of bus
$ $ $ P P P Multiple Busses • Simple way to increase bandwidth • use more than one bus • Can be static or dynamic assignment to busses • static • A->B always uses bus 0 • C->D always uses bus 1 • dynamic • arbitrate for a bus, like instruction dispatch to k identical CPU resources $ P
Crossbar • No bandwidth reduction • (except receiver at endoint) • Easy routing (on or offline) • Scales poorly • N2 area and delay • No locality
Hypercube • Arrange N=2n nodes in n-dimensional cube • At most n hops from source to sink • N = log2(N) • High bisection bandwidth • good for traffic (but can you use it?) • bad for cost [O(n2)] • Exploit locality • Node size grows • as log(N) [IO] • Maybe log2(N) [xbar between dimensions]
Multistage • Unroll hypercube vertices so log(N), constant size switches per hypercube node • solve node growth problem • lose locality • similar good/bad points for rest
Hypercube/Multistage Blocking • Minimum length multistage • many patterns cause bottlenecks • e.g.
CS184a: Day16 Beneš Network • 2log2(N)-1 stages (switches in path) • Made of N/2 22 switchpoints [4 sw] • 4Nlog2(N) total switches • Compute route in O(N log(N)) time • Routes all permutations
Online Hypercube Blocking • If routing offline, can calculate Benes-like route • Online, don’t have time, or global view • Observation: only a few, canonically bad patterns • Solution: Route to random intermediate • then route from there to destination • …turns worst-case into average case • at the expense of locality
K-ary N-cube • Alternate reduction from hypercube • restrict to N<log(Nodes) dimensional structure • allow more than 2 ordinates in each dimension • E.g. mesh (2-cube), 3D-mesh (3-cube) • Matches with physical world structure • Bounds degree at node • Has Locality • Even more bottleneck potentials • make channels wider (CS184a:Day 17)
Torus • Wrap around n-cube ends • 2-cube cylinder • 3-cube donut • Cuts worst-case distances in half • Can be laid-out reasonable efficiently • maybe 2x cost in channel width?
Fat-Tree • Saw that communications typically has locality (CS184a) • Modeled recursive bisection/Rent’s Rule • Leiserson showed Fat-Tree was (area, volume) universal • w/in log(N) the area of any other structure • exploit physical space limitations wiring in {2,3}-dimensions
MoT/Express Cube(Mesh with Bypass) • Large machine in 2 or 3 D mesh • routes must go through square/cube root switches • vs. log(N) in fat-tree, hypercube, MIN • Saw practically can go further than one hop on wire… • Add long-wire bypass paths
Issues/Axes • Throughput of Communication relative to data rate of media • Single point-to-point link consume media BW? • Can share links between multiple comm streams? • What is the sharing factor? • Binding time/Predictability of Interconnect • Pre-fab • Before communication then use for long time • Cycle-by-cycle • Network latency vs. persistence of communication • Comm link persistence
Axes Sharefactor (Media Rate/App. Rate) Persistence Predictability Net Latency
Hardwired • Direct, fixed wire between two points • E.g. Conventional gate-array, std. cell • Efficient when: • know communication a priori • fixed or limited function systems • high load of fixed communication • often control in general-purpose systems • links carry high throughput traffic continually between fixed points
Configurable • Offline, lock down persistent route. • E.g. FPGAs • Efficient when: • link carries high throughput traffic • (loaded usefully near capacity) • traffic patterns change • on timescale >> data transmission
Time-Switched • Statically scheduled, wire/switch sharing • E.g. TDMA, NuMesh, TSFPGA • Efficient when: • thruput per channel < thruput capacity of wires and switches • traffic patterns change • on timescale >> data transmission
Axes Time Mux Sharefactor (Media Rate/App. Rate) Predictability
Self-Route, Circuit-Switched • Dynamic arbitration/allocation, lock down routes • E.g. METRO/RN1 • Efficient when: • instantaneous communication bandwidth is high (consume channel) • lifetime of comm. > delay through network • communication pattern unpredictable • rapid connection setup important
Axes Phone; Videoconf; Cable Circuit Switch Sharefactor (Media Rate/App. Rate) Persistence Circuit Switch Net Latency Predictability
Self-Route, Store-and-Forward, Packet Switched • Dynamic arbitration, packetized data • Get entire packet before sending to next node • E.g. nCube, early Internet routers • Efficient when: • lifetime of comm < delay through net • communication pattern unpredictable • can provide buffer/consumption guarantees • packets small
Self-Route, Virtual Cut Through • Dynamic arbitration, packetized data • Start forwarding to next node as soon as have header • Don’t pay full latency of storing packet • Keep space to buffer entire packet if necessary • Efficient when: • lifetime of comm < delay through net • communication pattern unpredictable • can provide buffer/consumption guarantees • packets small
Virtual Cut Through Three words from same packet
Self-Route, Wormhole Packet-Switched • Dynamic arbitration, packetized data • E.g. Caltech MRC, Modern Internet Routers • Efficient when: • lifetime of comm < delay through net • communication pattern unpredictable • can provide buffer/consumption guarantees • message > buffer length • allow variable (? Long) sized messages
Wormhole Single Packet spread through net when not stalled
Wormhole Single Packet spread through net when stalled.
Phone; Videoconf; Cable Circuit Switch Persistence Packet Switch IP Packet SMS message Axes Packet Switch Time Mux Sharefactor (Media Rate/App. Rate) Circuit Switch Config urable Predictability Net Latency
Time Mux Dynamic Costs: Area • Area • switch (1-1.5Kl2/ switch) • larger with pipeline (4Kl2) and rebuffer • state (SRAM bit = 1.2Kl2/ bit) • multiple in time-switched cases • arbitrartion/decision making • usually dominates above • buffering (SRAM cell per buffer) • can dominate
Area (queue rough approx; you will refine)
Costs: Latency • Time single path • make decisions • round-trip flow-control • Time contention/traffic • blocking in buffers • quality of decision • pick wrong path • have stale data
Intermediate Approach • For large # of predictable patterns • switching memory may dominate allocation area • area of routed case < time-switched • [e.g. Large Cycles] • Get offline, global planning advantage • by source routing • source specifies offline determined route path • offline plan avoids contention
Offline vs. Online • If know patterns in advance • offline cheaper • no arbitration (area, time) • no buffering • use more global data • better results • As becomes less predictable • benefit to online routing
[example from Li and McKinley, IEEE Computer v26n2, 1993] Deadlock • Possible to introduce deadlock • Consider wormhole routed mesh
Dimension Order Routing • Simple (early Caltech) solution • order dimensions • force complete routing in lower dimensions before route in next higher dimension
[example from Li and McKinley, IEEE Computer v26n2, 1993] Dimension Ordered Routing • Route Y, then Route X
Dimension Order Routing • Avoids cycles in channel graph • Limits routing freedom • Can cause artificial congestion • consider • (0,0) to (3,3) • (1,0) to (3,2) • (2,0) to (3,1) • [There is a rich literature on how to do better]
Turn Model • Problem is cycles • Selectively disallow turns to break cycles • 2D Mesh West-First Routing